BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Data Compression with Relative Entropy Coding - Gergely Flamich\, 
 University of Cambridge
DTSTART:20250305T140000Z
DTEND:20250305T150000Z
UID:TALK219904@talks.cam.ac.uk
CONTACT:Prof. Ramji Venkataramanan
DESCRIPTION:In lossy source coding\, one aims to encode data using as few 
 bits as possible while ensuring we can approximately recover the data from
  the code with as little distortion as possible. Typically\, lossy source 
 coding algorithms comprise two steps: 1) they discard information by quant
 ising the source\, and 2) they encode the quantised representations lossle
 ssly with entropy coding algorithms such as arithmetic or Huffman coding. 
 Unfortunately\, this approach makes combining lossy source coding with mod
 ern machine learning approaches challenging since quantisation is a non-di
 fferentiable operation. However\, quantisation is not the only mechanism w
 e could use to discard information. Instead\, we could perturb it (e.g. wi
 th Gaussian noise) and encode the perturbed representation\; this is the e
 ssence of relative entropy coding. \n\nThis talk provides an introduction 
 to relative entropy coding. I begin by discussing the formal definition of
  the problem: what should "encoding a perturbed version of the data" even 
 mean? Then\, I provide an overview of its applications\, such as how we ca
 n use it for learned data compression\, low-rate compression with realism 
 constraints and differential privacy. Next\, I describe a relative entropy
  coding algorithm I developed: greedy Poisson rejection sampling (GPRS). T
 his involves a somewhat unorthodox use of Poisson processes: they form the
  basis of the algorithm's construction! Finally\, I discuss recent results
  on the fundamental limits of the communication and computational complexi
 ty of relative entropy coding algorithms and highlight some interesting op
 en problems and future research directions.
LOCATION:MR5\, CMS Pavilion A
END:VEVENT
END:VCALENDAR